guide
How to write AI video prompts that work
A text-to-video prompt is a short brief for a camera crew that does not exist. The model fills every gap you leave with a guess, and you pay for each guess. This guide gives you a five-part structure you can reuse for any model.
Prompt anatomy: click a part to see it in the example
Tracking shot following a red sports car as it drifts through a wet coastal corner, a police car close behind. Palm trees and pink neon signs line the road. Neon-lit night, tense mood, engine roar and tyre squeal.
Who or what is on screen, in concrete terms. “A red sports car” beats “a cool car”.
The five parts of a good prompt
- Subject: who or what is on screen, in concrete terms.
- Action: one thing the subject does.
- Setting: where it happens and what is around it.
- Camera: the shot type and movement.
- Look: lighting, mood and, if the model makes sound, what we hear.
Everything else is optional. If you only have time to get two parts right, make them the action and the camera, because those are the parts models most often get wrong when left to guess.
Before and after
Weak: “A cool car chase at night, cinematic.” It names no subject, no place, no camera and no light, so the model invents all four.
Stronger: “Tracking shot following a red sports car as it drifts through a wet coastal corner, a police car close behind. Palm trees and pink neon signs line the road. Neon-lit night, tense mood, engine roar and tyre squeal.” Every part of the formula is present, and nothing is left to luck.
Keep it to one action
Models handle short clips, often a few seconds long. A prompt that asks for a drift, a crash, a jump and a crowd reaction will usually deliver a blurred version of all four. Write one clip per action and join them in an editor.
Name the camera
- A tracking shot follows the subject and suggests speed.
- A slow push-in adds tension to a face or an object.
- A low-angle dolly makes vehicles and people look powerful.
- A handheld look feels documentary and nostalgic.
- A drone-style rise reveals the setting at the end of a clip.
Camera words steer motion more reliably than adjectives such as “epic”.
Say where the light comes from
“Golden hour sun”, “neon-lit night” and “soft overcast daylight” each set colour and mood in two or three words. Pick one lighting idea per prompt, because mixing several confuses the result.
Avoid real names
Video platforms often refuse prompts that name real people, brands, games or copyrighted characters, and the results can raise rights questions. Describe the look you want instead of naming the source.
Iterate one detail at a time
Generate a short test clip first. If it is wrong, change a single part of the prompt and try again. If you change five things at once you will not learn which change helped. Our budgeting guide explains why test clips are cheaper than full-length retakes.
Build the prompt faster
The free prompt generator on this site arranges these five parts for Veo, Sora, Kling and Runway, so you only fill in the details.
A worked example, part by part
Say you want a short clip of a coffee shop. Start with the subject: a barista in a green apron. Add the action: she pours milk into a cup in one slow stream. Add the setting: a small café with brick walls and morning light through the window. Add the camera: a slow push-in from the counter. Add the look: warm, soft, quiet, with the sound of a milk steamer. Put together, it reads as one clear shot a model can follow.
Notice what is missing: the story. A single clip does not need one. If you want a story, plan it as several clips and treat each prompt as one shot in a sequence.
Reusing a prompt across clips
If you need several clips that look like they belong together, keep the look and the camera style the same and change only the subject and action. A consistent lighting phrase, such as “soft overcast daylight”, ties clips together better than any single detail.
Frequently asked questions
How long should an AI video prompt be?
Most models do well with one to four sentences, roughly 40 to 120 words. Very short prompts leave too much to chance, and very long ones can dilute what matters.
Do I need a different prompt for each model?
The five parts stay the same. How you arrange them can differ: some models like a flowing paragraph, others like short labelled lines or comma-separated descriptors. Test a short clip on the model you use.
Why does the model ignore part of my prompt?
Models weigh some words more than others and cannot follow every instruction at once. If a detail keeps getting dropped, move it earlier in the prompt, simplify the rest, or give it its own clip.
Should I write prompts in full sentences?
Either works. Some models respond well to flowing sentences and others to short descriptors. Pick one style, test it, and stay consistent so you can compare results.
Keep reading
Some links on this site are affiliate links; see the disclosure. Platform rules and limits change, so confirm them on the platform’s own help pages.